rcomphierarc {actuar}

Simulation from Compound Hierarchical Models

Description

Simulate data for insurance applications allowing hierarchical structures and separate models for the frequency and severity of claims distributions.

rcomphierarc is an alias for simul.

Usage

rcomphierarc(nodes, model.freq = NULL, model.sev = NULL, weights = NULL)

## S3 method for class 'portfolio'
print(x, ...)

Arguments

nodes

a vector or a named list giving the number of "nodes" at each level in the hierarchy of the portfolio. The nodes are listed from top (portfolio) to bottom (usually the years of experience).

model.freq

a named vector of expressions specifying the frequency of claims model (see Details); if NULL, only claim amounts are simulated.

model.sev

a named vector of expressions specifying the severity of claims model (see Details); if NULL, only claim numbers are simulated.

weights

a vector of weights.

x

a portfolio object.

...

potential further arguments required by generic.

Details

The order and the names of the elements in nodes, model.freq and model.sev must match. At least one of model.freq and model.sev must be non NULL.

nodes may be a basic vector, named or not, for non hierarchical models. The rule above still applies, so model.freq and model.sev should not be named if nodes is not. However, for non hierarchical models, rcompound is faster and has a simpler interface.

nodes specifies the hierarchical layout of the portfolio. Each element of the list is a vector of the number of nodes at a given level. Vectors are recycled as necessary.

model.freq and model.sev specify the simulation models for claim numbers and claim amounts, respectively. A model is expressed in a semi-symbolic fashion using an object of mode expression. Each element of the object must be named and should be a complete call to a random number generation function, with the number of variates omitted. Hierarchical (or mixtures of) models are achieved by replacing one or more parameters of a distribution at a given level by any combination of the names of the levels above. If no mixing is to take place at a level, the model for this level can be NULL.

The argument of the random number generation functions for the number of variates to simulate must be named n.

Weights will be used wherever the name "weights" appears in a model. It is the user's responsibility to ensure that the length of weights will match the number of nodes when weights are to be used. Normally, there should be one weight per node at the lowest level of the model.

Data is generated in lexicographic order, that is by row in the output matrix.

Value

An object of class "portfolio". A print method for this class displays the models used in the simulation as well as the frequency of claims for each year and entity in the portfolio.

An object of class "portfolio" is a list containing the following components:

data

a two dimension list where each element is a vector of claim amounts;

weights

the vector of weights given in argument reshaped as a matrix matching element data, or NULL;

classification

a matrix of integers where each row is a unique set of subscripts identifying an entity in the portfolio (e.g. integers i, j and k for data X_{ijkt});

nodes

the nodes argument, appropriately recycled;

model.freq

the frequency model as given in argument;

model.sev

the severity model as given in argument.

It is recommended to manipulate objects of class "portfolio" by means of the corresponding methods of functions aggregate, frequency and severity.

Author(s)

Vincent Goulet [e-mail], Sébastien Auclair and Louis-Philippe Pouliot

References

Goulet, V. and Pouliot, L.-P. (2008), Simulation of compound hierarchical models in R, North American Actuarial Journal 12, 401–412.

See Also

rcomphierarc.summaries for the functions to create the matrices of aggregate claim amounts, frequencies and individual claim amounts.

rcompound for a simpler and much faster way to generate variates from standard, non hierarchical, compound models.

Examples

## Two level (contracts and years) portfolio with frequency model
## Nit|Theta_i ~ Poisson(Theta_i), Theta_i ~ Gamma(2, 3) and severity
## model X ~ Lognormal(5, 1)
rcomphierarc(nodes = list(contract = 10, year = 5),
             model.freq = expression(contract = rgamma(2, 3),
                                     year = rpois(contract)),
             model.sev = expression(contract = NULL,
                                    year = rlnorm(5, 1)))

## Model with weights and mixtures for both frequency and severity
## models
nodes <- list(entity = 8, year = c(5, 4, 4, 5, 3, 5, 4, 5))
mf <- expression(entity = rgamma(2, 3),
                 year = rpois(weights * entity))
ms <- expression(entity = rnorm(5, 1),
                 year = rlnorm(entity, 1))
wit <- sample(2:10, 35, replace = TRUE)
pf <- rcomphierarc(nodes, mf, ms, wit)
pf 				# print method
weights(pf)			# extraction of weights
aggregate(pf)[, -1]/weights(pf)[, -1] # ratios

## Four level hierarchical model for frequency only
nodes <- list(sector = 3, unit = c(3, 4),
              employer = c(3, 4, 3, 4, 2, 3, 4), year = 5)
mf <- expression(sector = rexp(1),
                 unit = rexp(sector),
                 employer = rgamma(unit, 1),
                 year = rpois(employer))
pf <- rcomphierarc(nodes, mf, NULL)
pf 				# print method
aggregate(pf) 			# aggregate claim amounts
frequency(pf)  			# frequencies
severity(pf)			# individual claim amounts

## Standard, non hierarchical, compound model with simplified
## syntax (function rcompound() is much faster for such cases)
rcomphierarc(10,
             model.freq = expression(rpois(2)),
             model.sev = expression(rgamma(2, 3)))